Pith. sign in

REVIEW 4 major objections 5 minor 38 references

From Sharpness to Better Generalization for Speech Deepfake Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Sharpness is proposed as a theoretical predictor of how well speech deepfake detectors generalize to unseen data.

desk verdict Useful empirical study with a real diagnostic corner, but the abstract oversells it and the sharpness-EER correlation has scale and grouping confounds that need addressing. read the letter →

arxiv 2506.11532 v1 pith:6CNFGZ4Z submitted 2025-06-13 eess.AS cs.SD

Wen Huang , Xuechen Liu , Xin Wang , Junichi Yamagishi , Yanmin Qian This is my paper
classification eess.AScs.SD
keywords speechdeepfakedetectiongeneralizationsharpnesssharpness-awareminimizationdomainshiftequalerrorrateself-supervisedlearningASVspoof
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Speech deepfake detectors are usually judged by equal error rate on unseen data, without a way to explain why one system generalizes better than another. This paper proposes that sharpness—how much a model's loss changes under small perturbations of its weights—can serve as that explanation. It reports that sharpness rises when test conditions shift away from training conditions, and that across 24 systems sharpness is positively correlated with EER on most out-of-domain test sets, most strongly on in-the-wild recordings (Spearman rank correlation 0.87). It further shows that Sharpness-Aware Minimization reduces sharpness and generally lowers EER while making results more stable across random seeds. If the claim holds, sharpness gives detector developers a measurable, theoretically motivated target to optimize.

What carries the argument

$$s(w,S) = \max_{\|\epsilon\|_2 \le \rho} \frac{1}{|S|} \sum_{i:(x_i,y_i)\in S} \bigl(\ell_i(w+\epsilon) - \ell_i(w)\bigr),$$ computed per batch and averaged. The paper pairs this measure with Sharpness-Aware Minimization, whose objective replaces the training loss by its maximum over a $\rho$-bounded perturbation neighborhood, approximated by a first-order gradient step. The fixed radius $\rho=0.05$ and batch size $m=32$ are kept constant when sharpness is compared across 24 systems; this measure is what carries the correlation analysis and the claim that SAM flattens the landscape.

What would settle it

A direct check would be to recompute the paper's 24-system sharpness–EER correlations with $\rho$ scaled per model, for example by parameter norm or layer-wise variance; if the reported correlations such as SRCC 0.87 on In-The-Wild disappear or reverse, the fixed-radius choice carried the result. A second check is to train a deliberately low-error model whose loss landscape is sharp; if its EER is low despite high sharpness, the claimed indicator fails in that setting.

Watch

Extended reading notes

Core claim

The paper's central claim is that sharpness is a usable theoretical indicator of generalization for speech deepfake detection, not just a heuristic. Concretely, the authors define m-sharpness on a test set as the maximum average loss increase under weight perturbations bounded by $\rho$, fix $\rho=0.05$ and batch size $m=32$, and show that sharper systems tend to have higher EER on unseen datasets, with statistically significant correlations in five of six out-of-domain settings. They also argue that minimizing sharpness through Sharpness-Aware Minimization, which optimizes the worst-case loss over a perturbation neighborhood, flattens the loss landscape and yields consistently better and more stable EER, with the largest gains on the most mismatched sets such as ASVspoof 2021 LA and ADD 2022.

Load-bearing premise

The central claim depends on the assumption that one fixed perturbation radius, $\rho=0.05$, gives fair and comparable sharpness values for models as different as AASIST and W2V-XLSR, and that the observed sharpness–EER correlation is not an artifact of that radius or of the Adam-versus-SAM grouping.

Editorial extensions

If this is right

  • A practitioner can measure sharpness on held-out data to predict which detectors will transfer, instead of relying on error-rate ranking alone.
  • SAM-trained models are a direct robustness recipe for deployment under language, channel, or attack mismatch, since gains are largest in those conditions.
  • Combining SSL pretraining with SAM is supported: even the flattest pretrained models usually improve further on out-of-domain test sets.
  • Sharpness can act as a diagnostic for which mismatch factors hurt a detector, since language, attack, and channel shifts raise it while speaker variability does not.
  • Training with SAM also reduces run-to-run variance, so the improvement is not just in the mean error rate.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the sharpness–error relationship is causal rather than merely predictive, tracking sharpness on a small validation set during training could serve for model selection without needing a labeled target-domain test set.
  • Editorial inference: the fixed-radius comparison may bundle flatness together with model scale; recomputing the correlations with $\rho$ scaled per parameter norm would test whether the ranking survives, an experiment the paper does not run.
  • Editorial inference: SAM's objective includes an explicit L2 term and per-model radius selection, so the improvement over Adam could partly come from those choices; an ablation holding weight decay fixed would isolate the perturbation term.
  • Editorial inference: the same sharpness measure could rank feature-level and data-level defenses as well as optimizers, giving a single scale on which to compare robustness interventions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes to use sharpness as a predictor of generalization for speech deepfake detection. It defines m-sharpness via Eq. (1), observes that sharpness increases under domain shifts for two models, and then applies Sharpness-Aware Minimization (SAM) to train AASIST and several wav2vec/XLS-R based systems on ASVspoof 2019 LA, evaluating on seven unseen benchmarks. The authors report that SAM lowers EER for many configurations (e.g., AASIST on 21LA: 8.10 to 4.62) and that sharpness correlates with EER across 24 systems, with the strongest correlations on ITW, 21LA, and ADD. The paper concludes that sharpness can serve as a theoretical indicator of generalization and that SAM improves robustness.

Significance. If the correlation claim were established with appropriate controls, the paper would be valuable: sharpness could serve as a cheap diagnostic for model selection and failure analysis, and SAM as a drop-in training method. The study is commendably broad, covering seven test sets, six model configurations, three seeds, and released code, and the concrete improvements on high-mismatch sets are practically interesting. However, the current analysis does not rule out that the sharpness-EER correlations are driven by architecture scale or by the Adam/SAM grouping, and the attribution of SAM gains to flatness is confounded with L2 regularization and per-model radius selection. These issues are fixable with additional analyses, so the work is a promising empirical contribution but not yet an established theoretical indicator.

major comments (4)
  1. [Section 2, Eq. (1); Section 4.1] The sharpness diagnostic in Eq. (1) uses a fixed perturbation radius ρ=0.05 for every model, whereas Section 4.1 selects SAM's radius per model from {0.05, 0.01, 0.005, 0.001}, with smaller models assigned larger values. Because Eq. (1) is not normalized by parameter norm, the same L2 radius produces very different relative perturbations for AASIST versus SSL encoders with much larger weight norms, so the lower sharpness of SSL models and the correlations in Table 2 may reflect architecture scale rather than a general sharpness-generalization relationship. Please report within-family correlations, within-optimizer correlations, or partial correlations controlling for parameter count, and include a sensitivity analysis over ρ.
  2. [Table 2 and Figure 3] The correlation coefficients in Table 2 are computed over all 24 systems pooled across Adam and SAM. Since SAM simultaneously reduces both EER and sharpness for most model-dataset pairs (e.g., AASIST on 21LA in Table 1), the between-optimizer separation can inflate Spearman and Pearson coefficients even if no within-optimizer trend exists. Figure 3 appears to show separate regression lines for the two groups, but the corresponding within-group coefficients are not reported. Please report correlations for Adam-only and SAM-only subsets, or include the optimizer as a covariate.
  3. [Section 4.2, Table 1] The text says SAM 'consistently improves generalization across nearly all models and datasets,' but Table 1 contains several clear regressions: W2V-Base+Linear on WF (30.34 to 36.95), W2V-Large+Linear on FOR (13.24 to 14.65) and ADD (30.97 to 37.31), W2V-XLSR+Linear on 21LA (1.32 to 1.87), and W2V-XLSR+AASIST on SC (27.50 to 29.93). Please quantify the fraction of pairwise improvements and either soften the abstract/conclusion claims or analyze what distinguishes the configurations where SAM hurts.
  4. [Eq. (2) and Section 4.1] The paper attributes SAM's performance gains to flatness, but the SAM objective in Eq. (2) also includes an explicit L2 term λ||w||^2, and the SAM radius is tuned per model in Section 4.1. Without ablating against standard Adam with equivalent L2 regularization, or evaluating sharpness at each model's own selected ρ, the causal link from reduced sharpness to improved EER is not isolated. Please add such an ablation or restrict the claim to a correlational one.
minor comments (5)
  1. [Section 4.1] There is a duplicated phrase 'in in 4.1' in the sentence describing the dataset and training configurations.
  2. [Table 1] The header 'W2V-Base+AASISTAdam' is missing a space between 'AASIST' and 'Adam'.
  3. [Figure 2] The loss-landscape visualization covers only one model-dataset pair, so the claim that SAM 'results in a noticeably flatter loss landscape' would be stronger if additional configurations were shown; as presented, the visualization is illustrative only.
  4. [Table 2] The p-values for the correlation coefficients are not adjusted for multiple comparisons across seven datasets and three metrics; please state whether the reported statistical significance survives a multiple testing correction.
  5. [Abstract and Conclusion] The phrase 'theoretical indicator' overstates what is currently an empirical correlation; consider using 'empirical indicator' or 'diagnostic' unless the requested confound controls change the strength of the evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the sharpness-EER relationship is empirical, and SAM's objective does not by construction force EER improvement.

full rationale

The paper's load-bearing claim is empirical rather than derivational: sharpness is measured on test sets using Eq. (1) (adopted from Foret et al. [21] and Andriushchenko & Flammarion [23]), EER is measured on unseen evaluation sets, and the association is assessed via PCC/SRCC/KTAU in Table 2. No parameter is fitted to the target EER, and the sharpness values are not defined in terms of EER. Although SAM's objective in Eq. (2) contains the same sharpness term as Eq. (1), minimizing sharpness does not logically guarantee lower EER; indeed Table 1 contains counterexamples (e.g., W2V-Base+Linear on WF: SAM 36.95 vs Adam 30.34; W2V-XLSR+AASIST on SC: SAM 29.93 vs Adam 27.50), so the correlation is not forced by construction. Citations to the authors' own prior work are used as datasets/benchmarks or standard external references, not as the justification for the sharpness-generalization relationship. Potential confounds (fixed rho=0.05 across differently scaled models, pooling Adam and SAM groups) are validity and generalizability concerns, not circularity. Therefore no specific circular step can be exhibited.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central empirical results rest on the sharpness definition, the chosen perturbation radii, and the comparability of sharpness across very different model scales. These are domain assumptions, not derived facts. The paper introduces no new entities; its 'theoretical' label is a high-level framing rather than a formal theory.

free parameters (4)
  • SAM perturbation bound rho (per model) = chosen from {0.05, 0.01, 0.005, 0.001}
    Hyperparameter for SAM; larger models get smaller rho. No cross-validation or sensitivity analysis. Central to the sharpness reduction claim.
  • Sharpness evaluation rho = 0.05
    Fixed perturbation bound for all sharpness computations (Section 2, Figure 1, Table 2). Scale of this value affects all correlations and sharpness comparisons across architectures.
  • Batch size m for sharpness = 32
    Used for m-sharpness; chosen without justification.
  • Class weights for weighted cross-entropy = 0.9 bona fide / 0.1 spoof
    Training choice that affects loss landscape and sharpness; no ablation reported.
assumptions (4)
  • domain assumption The sharpness definition of Eq. (1) is a meaningful proxy for generalization in SDD.
    The paper's entire diagnostic and correlation analysis assumes that test-set m-sharpness tracks generalization; no derivation or theory is provided.
  • standard math The first-order perturbation approximation of SAM (Eqs. 4-5) is accurate enough to reduce sharpness as defined in Eq. (1).
    Taken from Foret et al. [21]; the paper relies on this approximation without new justification.
  • domain assumption Sharpness values computed with a fixed rho are comparable across architectures with very different parameter scales (AASIST vs W2V-Base/XLSR).
    The paper compares sharpness across models of different sizes without weight-normalization; this comparability is essential to the model-level correlation.
  • domain assumption EER is the appropriate ground-truth generalization measure for SDD.
    Standard in the field; used as the y-axis for all correlation claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Sharpness to Better Generalization for Speech Deepfake Detection." pith.science (2026). https://pith.science/paper/6CNFGZ4Z

@misc{pith2026250611532,
  author       = {Pith},
  title        = {Pith review of: From Sharpness to Better Generalization for Speech Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6CNFGZ4Z}},
  note         = {Machine review of arXiv:2506.11532}
}
read the original abstract

Generalization remains a critical challenge in speech deepfake detection (SDD). While various approaches aim to improve robustness, generalization is typically assessed through performance metrics like equal error rate without a theoretical framework to explain model performance. This work investigates sharpness as a theoretical proxy for generalization in SDD. We analyze how sharpness responds to domain shifts and find it increases in unseen conditions, indicating higher model sensitivity. Based on this, we apply Sharpness-Aware Minimization (SAM) to reduce sharpness explicitly, leading to better and more stable performance across diverse unseen test sets. Furthermore, correlation analysis confirms a statistically significant relationship between sharpness and generalization in most test settings. These findings suggest that sharpness can serve as a theoretical indicator for generalization in SDD and that sharpness-aware training offers a promising strategy for improving robustness.

Figures

Figures reproduced from arXiv: 2506.11532 by the authors.

Figure 1
Figure 1. Sharpness values computed for two models, AASIST and W2V-Base+Linear, across different mismatch scenarios. Abbreviations for test sets: WF (WaveFake), SC (SpoofCeleb), 21LA (ASVspoof 2021 LA). sharpness and generalization by computing correlation metrics between sharpness and EER and find a statistically significant correlation in most out-of-distribution datasets. Overall, these findings suggest that sharpness can … view at source ↗
Figure 2
Figure 2. Visualization of the loss landscape for W2V￾Base+Linear trained with the Adam optimizer (left) and the SAM optimizer (right). The x-y axes represent scaling factors for two weight perturbation directions, with the original model at (0,0). The z-axis shows the corresponding loss values for the perturbed model. In SAM optimization, the perturbation bound ρ was chosen from {0.05, 0.01, 0.005, 0.001}, with smaller model… view at source ↗
Figure 3
Figure 3. Scatter plots of sharpness (x-axis) and EER (%) (y-axis) across seven datasets. Each subplot corresponds to one dataset, with square markers for Adam systems and cross markers for SAM systems. A linear regression trend line is included for each group to represent the relationship between the two variables. directions, and evaluated the loss on the 19LA evaluation set. The visualization shows that SAM results in a no… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 32 canonical work pages

  1. [1]

    Introduction Speech Deepfake Detection (SDD) aims to differentiate gen- uine from synthetic speech, which plays a vital role in com- bating audio-based misinformation, fraud, and security threats. However, generalization remains a major challenge, as mod- els trained on specific datasets often struggle when tested on unseen data, particularly under mismat...

    arXiv 2025

  2. [2]

    Given this characteristic, we aim to first explore whether sharpness can serve as a diagnos- tic tool to evaluate how domain mismatch impacts the model’s sensitivity and robustness

    Sharpness and Domain Mismatch Sharpness measures the sensitivity of the loss function to per- turbations in the model’s parameters, which is often associated with the model’s robustness [21]. Given this characteristic, we aim to first explore whether sharpness can serve as a diagnos- tic tool to evaluate how domain mismatch impacts the model’s sensitivity...

    work page 2019

  3. [3]

    Similar to Section 2, let S = (xi, yi)n i=1 be a set of la- beled data, and let ℓi(w) denote the loss evaluated at the point (xi, yi)

    Sharpness-Aware Minimization Sharpness-Aware Minimization is proposed in [21] to improve model generalization by simultaneously minimizing the loss value and its sharpness. Similar to Section 2, let S = (xi, yi)n i=1 be a set of la- beled data, and let ℓi(w) denote the loss evaluated at the point (xi, yi). For simplicity, we use LS(w) = 1/|S| Pn i=1 ℓi(w)...

    work page

  4. [4]

    Experimental Settings Datasets

    Experiments and Analysis 4.1. Experimental Settings Datasets. Systems were trained on the ASVspoof 2019 LA training set [24], which consists of 25k clean English utter- ances, including spoofed samples generated using 6 different voice conversion and text-to-speech attacks. The evaluation was conducted on eight unseen datasets representing diverse domain ...

    work page 2019

  5. [5]

    Our findings show that sharpness increases under domain shifts, indicating model sensitivity to unseen conditions

    Conclusion In this work, we examined sharpness as a theoretical proxy for generalization in speech deepfake detection. Our findings show that sharpness increases under domain shifts, indicating model sensitivity to unseen conditions. Our correlation anal- ysis confirmed its positive association with generalization, es- pecially in high-mismatch scenarios....

    work page

  6. [6]

    This study is partially supported by JST AIP Acceleration Research (JPMJCR24U3)

    Acknowledgements This work was conducted during the first author’s intern- ship at NII, Japan. This study is partially supported by JST AIP Acceleration Research (JPMJCR24U3). This work was also supported in part by China NSFC projects under Grants 62122050 and 62071288, in part by Shanghai Munici- pal Science and Technology Commission Project under Grant...

    work page

  7. [7]

    ASVspoof 2021: Accelerating progress in spoofed and deep- fake speech detection,

    J. Yamagishi, X. Wang, M. Todisco, M. Sahidullah, J. Patino, A. Nautsch, X. Liu, K. A. Lee, T. Kinnunen, N. Evans et al. , “ASVspoof 2021: Accelerating progress in spoofed and deep- fake speech detection,” in ASVspoof 2021 Workshop-Automatic Speaker Verification and Spoofing Coutermeasures Challenge , 2021

    work page 2021

  8. [8]

    Does audio deepfake detection generalize?

    N. M. M ¨uller, P. Czempin, F. Dieckmann, A. Froghyar, and K. B¨ottinger, “Does audio deepfake detection generalize?” Proc. ISCA Interspeech, 2022

    work page 2022

Show all 38 references
  1. [9]

    Raw- Boost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,

    H. Tak, M. Kamble, J. Patino, M. Todisco, and N. Evans, “Raw- Boost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,” in Proc. ICASSP. IEEE, 2022, pp. 6382–6386

    2022

  2. [10]

    SpecAugment: A simple data augmen- tation method for automatic speech recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “SpecAugment: A simple data augmen- tation method for automatic speech recognition,” arXiv preprint arXiv:1904.08779, 2019

    1904 arXiv

  3. [11]

    Spoofed training data for speech spoofing countermeasure can be efficiently created using neural vocoders,

    X. Wang and J. Yamagishi, “Spoofed training data for speech spoofing countermeasure can be efficiently created using neural vocoders,” in Proc. ICASSP. IEEE, 2023, pp. 1–5

    2023

  4. [12]

    Can large-scale vocoded spoofed data improve speech spoofing countermeasure with a self-supervised front end?

    ——, “Can large-scale vocoded spoofed data improve speech spoofing countermeasure with a self-supervised front end?” in Proc. ICASSP. IEEE, 2024, pp. 10 311–10 315

    2024

  5. [13]

    CPAUG: Refining copy-paste augmentation for speech anti-spoofing,

    L. Zhang, K. A. Lee, L. Zhang, L. Wang, and B. Niu, “CPAUG: Refining copy-paste augmentation for speech anti-spoofing,” in ICASSP 2024-2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 10 996–11 000

    2024

  6. [14]

    AASIST: Audio anti-spoofing using integrated spectro-temporal graph attention networks,

    J.-w. Jung, H.-S. Heo, H. Tak, H.-j. Shim, J. S. Chung, B.-J. Lee, H.-J. Yu, and N. Evans, “AASIST: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” in Proc. ICASSP. IEEE, 2022, pp. 6367–6371

    2022

  7. [15]

    Spoofing speech detection by modeling local spectro-temporal and long-term dependency,

    H. Wu, W. Guo, Z. Zhang, W. Zhao, S. Peng, and J. Zhang, “Spoofing speech detection by modeling local spectro-temporal and long-term dependency,” inProc. ISCA Interspeech, 2024, pp. 507–511

    2024

  8. [16]

    Temporal-channel modeling in multi-head self- attention for synthetic speech detection,

    D.-T. Truong, R. Tao, T. Nguyen, H.-T. Luong, K. A. Lee, and E. S. Chng, “Temporal-channel modeling in multi-head self- attention for synthetic speech detection,” in Proc. ISCA Inter- speech, 2024, pp. 537–541

    2024

  9. [17]

    wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,” Advances in neural information processing systems , vol. 33, pp. 12 449–12 460, 2020

    2020

  10. [18]

    XLS-R: Self-supervised cross-lingual speech rep- resentation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Conneau, and M. Auli, “XLS-R: Self-supervised cross-lingual speech rep- resentation learning at scale,” arXiv, vol. abs/2111.09296, 2021

    2021 arXiv

  11. [19]

    WavLM: Large-scale self- supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “WavLM: Large-scale self- supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  12. [20]

    One-class learning towards syn- thetic voice spoofing detection,

    Y . Zhang, F. Jiang, and Z. Duan, “One-class learning towards syn- thetic voice spoofing detection,” IEEE Signal Processing Letters, vol. 28, pp. 937–941, 2021

    2021

  13. [21]

    One-class learning with adap- tive centroid shift for audio deepfake detection,

    H. M. Kim, K. Jang, and H. Kim, “One-class learning with adap- tive centroid shift for audio deepfake detection,” in Proc. ISCA Interspeech, 2024, pp. 4853–4857

    2024

  14. [22]

    Generalizable audio deepfake detection via latent space refinement and augmen- tation,

    W. Huang, Y . Gu, Z. Wang, H. Zhu, and Y . Qian, “Generalizable audio deepfake detection via latent space refinement and augmen- tation,” in Proc. ICASSP, 2025, pp. 1–5

    2025

  15. [23]

    Tandem assessment of spoofing countermeasures and au- tomatic speaker verification: Fundamentals,

    T. Kinnunen, H. Delgado, N. Evans, K. A. Lee, V . Vestman, A. Nautsch, M. Todisco, X. Wang, M. Sahidullah, J. Yamagishi et al., “Tandem assessment of spoofing countermeasures and au- tomatic speaker verification: Fundamentals,” IEEE Trans. Audio, Speech, Language Process., vol...

    2020

  16. [24]

    t- EER: Parameter-free tandem evaluation of countermeasures and biometric comparators,

    T. H. Kinnunen, K. A. Lee, H. Tak, N. Evans, and A. Nautsch, “t- EER: Parameter-free tandem evaluation of countermeasures and biometric comparators,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  17. [25]

    On large-batch training for deep learning: Gen- eralization gap and sharp minima,

    N. S. Keskar, D. Mudigere, J. Nocedal, M. Smelyanskiy, and P. T. P. Tang, “On large-batch training for deep learning: Gen- eralization gap and sharp minima,” 2017

    2017

  18. [26]

    Fantastic generalization measures and where to find them,

    Y . Jiang, B. Neyshabur, H. Mobahi, D. Krishnan, and S. Bengio, “Fantastic generalization measures and where to find them,” 2020

    2020

  19. [27]

    Sharpness- aware minimization for efficiently improving generalization,

    P. Foret, A. Kleiner, H. Mobahi, and B. Neyshabur, “Sharpness- aware minimization for efficiently improving generalization,” in International Conference on Learning Representations, 2021

    2021

  20. [28]

    Multi-dataset co- training with sharpness-aware optimization for audio anti- spoofing,

    H.-j. Shim, J.-w. Jung, and T. Kinnunen, “Multi-dataset co- training with sharpness-aware optimization for audio anti- spoofing,” in Proc. ISCA Interspeech, 2023, pp. 3804–3808

    2023

  21. [29]

    Towards understanding sharpness-aware minimization,

    M. Andriushchenko and N. Flammarion, “Towards understanding sharpness-aware minimization,” in International Conference on Machine Learning. PMLR, 2022, pp. 639–668

    2022

  22. [30]

    ASVspoof 2019: A large-scale public database of synthe- sized, converted and replayed speech,

    X. Wang, J. Yamagishi, M. Todisco, H. Delgado, A. Nautsch, N. Evans, M. Sahidullah, V . Vestman, T. Kinnunen, K. A. Lee et al., “ASVspoof 2019: A large-scale public database of synthe- sized, converted and replayed speech,” Computer Speech & Lan- guage, vol. 64, p. 101114, 2020

    2019

  23. [31]

    FoR: A dataset for synthetic speech detection,

    R. Reimao and V . Tzerpos, “FoR: A dataset for synthetic speech detection,” in 2019 International Conference on Speech Technol- ogy and Human-Computer Dialogue (SpeD) . IEEE, 2019, pp. 1–10

    2019

  24. [32]

    WaveFake: A data set to facilitate audio deepfake detection,

    J. Frank and L. Sch ¨onherr, “WaveFake: A data set to facilitate audio deepfake detection,” in Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021

    2021

  25. [33]

    The LJ Speech dataset,

    K. Ito and L. Johnson, “The LJ Speech dataset,” https://keithito. com/LJ-Speech-Dataset/, 2017

    2017

  26. [34]

    JSUT corpus: Free large-scale japanese speech corpus for end-to-end speech synthe- sis,

    R. Sonobe, S. Takamichi, and H. Saruwatari, “JSUT corpus: Free large-scale japanese speech corpus for end-to-end speech synthe- sis,” arXiv preprint arXiv:1711.00354, 2017

    2017 arXiv

  27. [35]

    ADD 2022: the first audio deep synthesis detection challenge,

    J. Yi, R. Fu, J. Tao, S. Nie, H. Ma, C. Wang, T. Wang, Z. Tian, Y . Bai, C. Fanet al., “ADD 2022: the first audio deep synthesis detection challenge,” in Proc. ICASSP. IEEE, 2022, pp. 9216– 9220

    2022

  28. [36]

    SpoofCeleb: Speech deepfake detection and SASV in the wild,

    J.-w. Jung, Y . Wu, X. Wang, J.-H. Kim, S. Maiti, Y . Matsunaga, H.-j. Shim, J. Tian, N. Evans, J. S. Chung et al., “SpoofCeleb: Speech deepfake detection and SASV in the wild,” IEEE Open Journal of Signal Processing, 2025

    2025

  29. [37]

    Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,

    H. Tak, M. Todisco, X. Wang, J.-w. Jung, J. Yamagishi, and N. Evans, “Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,” in The Speaker and Language Recognition Workshop, 2022

    2022

  30. [38]

    Visualizing the loss landscape of neural nets,

    H. Li, Z. Xu, G. Taylor, C. Studer, and T. Goldstein, “Visualizing the loss landscape of neural nets,”Advances in neural information processing systems, vol. 31, 2018

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.